US20240296614
2024-09-05
Physics
G06T13/80
Systems and methods for animating portraits in real-time are described, focusing on creating photorealistic animations from a single target image. The process begins by receiving scenario data that details the movements of a source head, along with a target image containing a second head and its background. The technology determines the necessary 2D deformations of the second head based on the movements captured from the first head.
The method involves applying the computed 2D deformations to the target image, resulting in an output frame that shows the second head mimicking the movements of the first head. Additionally, gaps between the displaced second head and its background are filled using a neural network designed for background prediction, enhancing the overall realism of the animation.
This technology can be implemented on various devices, including mobile phones and computers, allowing for real-time processing without needing extensive server resources. The approach is designed to work efficiently on standard mobile devices, making it accessible for widespread use in applications such as entertainment and virtual reality.
Unlike traditional methods that often require multiple images or videos to achieve realistic results, this innovation only requires a single target image. It significantly reduces computational time while maintaining high-quality outputs. By utilizing 3D modeling for generating 2D deformations, it streamlines the animation process compared to more complex techniques that depend on accurate segmentation and texture mapping.
The system allows users to create scenarios by selecting specific facial expressions and movements they wish to see animated on the target face. This user-friendly approach provides flexibility and creativity in generating animations, enabling a range of expressions such as frowns or smiles to be applied seamlessly to the portrait in real-time.